Multi-task Pedestrian Attribute Recognition Based on Machine Attention in Surveillance Scenarios

As a high-level semantic feature, pedestrian attributes have been widely used in person re-identification(Re-ID) and video analysis due to their robustness to lighting and posture changes. To improve the accuracy of pedestrian attribute recognition (PAR) in monitoring scene, the latest artificial intelligence (AI) explanatory research was referred, and the perspective of the machine attention was considered from. Considering comprehensively the mutual exclusion and dependence of attributes, a multi-task attribute recognition method limiting the attention of the neural network was proposed. First, according to the different areas of the neural network for each attribute, the pedestrian attribute recognition was divided into multiple sub-tasks. Then through the design of the end-to-end network model and the auxiliary classification loss function, the shared information among the subtasks was controlled, and the internal attributes of the subtasks to compete with each other was encouraged. On the contrary, the attributes between different subtasks were improved each other. Finally, the pedestrian attributes were predicted by fusing the information of each subtask. Through experiments, the proposed method has achieved the best accuracy on two public datasets in monitoring scenarios, which proves the effectiveness of the method.